Gitea evaluates a job's if-gate only once a runner with a matching
label claims it, so windows-latest cells wait forever and pin every
run (and the badge) at Waiting. The README table now says plainly
which rows CI covers (the Linux ones) and which are field-tested.
uv's cache moves into the workspace so it shares a filesystem with
.venv and hardlinks instead of double-storing ~1.3 GB per job (the
'Failed to hardlink, falling back to full copy' warning was the tell
before ENOSPC), and both matrices run max-parallel 1 so the host
holds one dependency set at a time instead of six.
github.ref reaches the shell via env instead of ${{ }} interpolation
- a crafted PR branch name could otherwise inject commands into the
runner - and ref/sha formats are validated before git sees them.
act_runner injects no node runtime into job containers the way
GitHub's runners do, so the JS checkout action dies in plain distro
images with 'node: executable file not found in $PATH'. A git
init/fetch/checkout needs only git, which each distro setup step now
installs before checkout runs.
Gitea Actions matrix: ubuntu-latest (py3.11/3.13) plus Debian 12,
Ubuntu 24.04, Fedora and Arch containers, each installing the system
libraries PySide6 wheels dlopen - all four validated locally in
Docker (Debian 141/141, QtWidgets import probe green everywhere).
Windows/macOS and a NixOS-FHS job (tools/ci-fhs.nix, the library set
from docs/NIXOS.md) are gated to github.com runners so they skip
instead of queueing forever here. README gets a Platform support
table with the workflow badge; KiCad-in-the-loop stays field-tested,
CI covers the solver and fallback suite.
QtCore's libraries (glib, icu) can load while QtWidgets/QtGui still
miss libGL - matplotlib's qt backend imports all three, so the old
probe could promise QtAgg and lose even the error figure at
switch_backend, one library layer deeper than the NixOS fix. The
regression test's mock now mirrors the real failure (bare package and
QtCore import fine, GUI modules raise): both the bare-import and the
QtCore-probe variants now fail it, verified by mutation. Also drop two
redundant f-prefixes on placeholder-less strings in report.py.
On OK the dialog closed and nothing appeared until the figures did,
which on a real board is minutes of looking like the plugin did
nothing. Put a small always-on-top window up for that stretch: the
stage now running, elapsed seconds, and Cancel.
Qt only repaints while the event loop runs and the solve owns the
thread, so the window pumps events itself - from inside the CG/AMG
iteration callback, which is where the time actually goes. That is
also where Cancel is noticed. The state is module-level because the
tick happens several frames deep in scipy/pyamg, and threading a
handle through those signatures for a progress bar is not worth it;
it stays inert until start(), so the standalone runner and the tests
are unaffected.
The window covers the figure work too, not just the solve: laying out
labels and writing four PNGs at full DPI is seconds on a real board -
10-15 of them on a large one - and closing before that left the same
silent gap one step later.
Deriving the body from the tag annotation does not survive CI: the
release action does not fall back to it (v1.2.0 published empty), and
checkout leaves the tag lightweight, so %(contents) yields the commit
message instead - which is what v1.2.1 first published. Keep the notes
at docs/release-notes/v<version>.md, pass that as body_path, and fail
the run when it is missing.
checkout fetches with --no-tags unless fetch-tags is set, at any
fetch-depth, so the notes step read no tag and wrote an empty file -
v1.2.1 published with empty notes despite the previous commit. Fail
the step rather than publish empty notes a third time.
akkuman/gitea-release-action takes body/body_path but does not fall
back to the tag annotation, so v1.2.0 published with empty notes.
Write the tag body to a file and pass it; fetch-depth 0 so the
annotated tag object is present to read it from.
kipy's Board.remove_items discards the DeleteItemsResponse, and the
proto warns the overall status may read OK even when nothing was
deleted - a locked image comes back IDS_IMMUTABLE. remove_overlays
went through the proto layer already for create; do the same for
delete, so a locked overlay is reported instead of silently surviving
while a second image is stacked on top of it.
Slots that a narrower run does not write kept the previous solve's
heatmap and read as current; clear them. The whole push is now one
commit, so a single undo reverts it rather than one layer of it.
_pad_polygons probed F.Cu before B.Cu regardless of which side the
joint protrudes from, so a pad sized differently per copper layer had
its solder coat measured from the wrong face; probe the solder side
first. Pads on no single copper layer are now noted rather than
silently skipped.
_fail could report nothing at all: before the output directory exists
no PNG is written, and with no GUI toolkit plots only opens saved
PNGs - the broken-plugin-environment case the docstring promises to
cover. Fall back to the temp dir, and never let reporting mask the
fault. Frequency input now keeps its specific rejection reason, as
the other numeric fields do.
- pyproject.toml + uv.lock: uv-managed dev environment (uv sync /
uv run pytest). requirements.txt stays: KiCad builds the plugin's
runtime venv from it and the PCM zip packages it.
- README: em-dash and run-on cleanup (34 -> 22, the rest deliberate),
split the 120-word THT solder-joint sentence, fix the documented
ADAPTIVE_MAX_CELL_UM value (2 mm -> 1 mm, config has 1000 um), fix
the *Packaging / publishing* cross-reference, dev sections now use
uv sync / uv run.
- LLM disclaimer: "most commits" carry the trailer (32 of 39), figures
claim now excepts the hand-drawn hole cross-section, reference the
UT3513+ measured-vs-computed validation.
- .gitignore: local AI-tooling artifacts; deploy scripts exclude
pyproject.toml/uv.lock; error-figure title punctuation aligned with
the other window titles.
- adaptive: barrel links could attach to coarse leaves (up to 1 mm),
making the whole leaf equipotential and deleting the local spreading
resistance - via-field results read up to ~13% low. Barrel attachment
cells are now pinned into the keep-fine set before the quadtree is
built; the guard ring grades around them.
- skin/dialog: the decimal-comma rewrite turned '1,500' into 1.5, a
silent 1000x error in frequency, test current or cell size. Ambiguous
comma patterns (thousands separators, multiple commas, mixed with a
dot) now raise with a message; a real decimal comma ('1,5') still
parses.
- dialog: Selection.adaptive default now matches the documented
on-by-default; the adaptive accuracy claim is aligned to the measured
0.03% quoted in README and config.
Swarm-review bug fixes: empty-layer crashes on the adaptive grid,
teardrop fills included in the copper model, exact track bbox, correct
AC current-density normalization, true 1D-trace current density.
Mutable tags on third-party actions could be repointed at code that
runs with the repo token; pin checkout, upload-artifact and
gitea-release-action to their current commits.
Every push to main builds the addon zip + registry metadata as a CI
artifact; pushing a v<version> tag (checked against metadata.json)
creates a release with both attached, matching the download_url
convention build_package.py already writes.
Follow the swarm-review fixes: chain cells display the true in-trace
current density, teardrop zones conduct, and AC current-density maps
are referenced to the skin-reduced equivalent thickness.
- adaptive: skip layers with zero quadtree leaves in the connectivity
restriction and mesh-boundary loops (IndexError on boards where a
selected layer has no copper)
- board_io: accept ZT_TEARDROP zones as conducting copper (KiCad types
teardrop fills ZT_TEARDROP, never ZT_COPPER, so they were dropped)
- geometry: copper_bbox uses the exact stroke bbox (centerline extrema
+ half width) instead of a 100 um chord tessellation that could
undershoot arc/cap extrema past the raster guard margin
- solver/adaptive: reference |J| to the conduction-equivalent thickness
sigma*rho in every branch (the uniform branch used geometric t, so AC
plots changed scale ~rs_ratio depending on unrelated per-cell maps)
- solver/adaptive/raster: chain cells no longer show phantom sheet-face
currents; store dl per chain link and overlay the true 1D density
|dV|/(rho*dl) (exact at any frequency: AC scaling of link conductance
and cross-section cancels)
- test_quadtree: compare edge lists pair-for-pair (the independent
column sort destroyed endpoint association)
grabowski